Property Valuation Accuracy: AI vs Manual Appraisal

The question acquisition teams ask us most often is not whether automated valuation models are useful. They have been using CoreLogic's AVM tools for years. The question is where the automated estimate is reliable enough to drive an offer anchor, and where it breaks down in ways that create real risk.

We ran a structured comparison across 80 residential and strata commercial properties in the Sydney metro area over a 6-month period in late 2025, comparing automated valuation outputs against certified appraisals from qualified valuers. The results were instructive, and they were not uniformly favourable to automated methods.

The Comparison Methodology

All 80 properties were within the Greater Sydney metropolitan area, covering a range of asset types: freestanding houses, terrace dwellings, strata units, and strata-titled commercial lots. Each property had a certified valuation report from a registered property valuer, and an automated estimate produced by our internal model within 7 days of the valuation date.

We measured error as the absolute percentage difference between the automated estimate and the certified value, and segmented results by asset type, suburb tier, and whether the property had transacted publicly within the prior 24 months.

Where Automated Valuation Performs Well

For properties with high comparability, the automated model performed with a median absolute percentage error below 4%. High comparability means a property type that transacts regularly within a well-defined geographic cluster, where recent sale data provides dense comparison points. Inner-ring strata units in suburbs like Pyrmont, Zetland, and Surry Hills fall into this category. There are enough comparable sales within 500 metres and 18 months that a hedonic pricing model can weight substitutes reliably.

Freestanding dwellings in established suburbs with regular turnover also performed well. Properties in the $1.2m to $2.8m range in suburbs like Marrickville and Newtown produced median errors in the 3.5% to 5% range. For initial deal screening purposes, a 5% error range on a $2m property, a $100,000 band, is operationally acceptable when the alternative is spending four hours on a preliminary manual assessment.

The speed advantage of automated valuation is most pronounced in deal flow triage. When an acquisition team receives 15 new signal alerts in a week, automated valuation provides a rapid estimate that determines whether each property merits deeper investigation. That triage function does not require sub-2% accuracy.

Where Automated Valuation Breaks Down

The failure modes were concentrated in three property categories.

First, properties with idiosyncratic physical attributes. A period terrace with a rear warehouse conversion on a rear-facing site in Darlinghurst produced an automated estimate 22% below the certified value. The model had no comparable sales that matched the dual-use character of the property, and the substitution logic defaulted to the lower price band of unrenovated terraces in the same street. A human valuer assessed the conversion quality and the tenancy income separately and arrived at a materially higher figure.

Second, strata commercial with short or expired leases. The automated model treats short leases as a negative adjustment but cannot assess the probability of renewal based on tenant quality, covenant strength, or the specific terms of the lease. A ground-floor retail unit in Glebe with a 14-month remaining lease had an automated estimate 17% below the certified value, where the certified valuer had assessed the tenant as a strong renewal candidate and applied an income capitalisation approach accordingly.

Third, properties in suburbs with thin transaction volumes. In outer suburban areas with fewer than 8 comparable sales in a 12-month radius, the model's confidence degrades substantially. The comparable set is too sparse to weight reliably, and the median error in this segment reached 11%.

The Practical Framework for Using Both

The takeaway from this comparison is not that automated valuation should replace manual appraisal. It is that automated and manual valuation serve different functions in the acquisition workflow, and conflating them creates risk at the wrong decision points.

Use automated valuation for: initial screening, setting offer range anchors on high-comparability properties, portfolio monitoring of assets already held, and quick-turn assessment when a deal needs a preliminary view in under 24 hours.

Require a certified appraisal for: offer finalisation on any asset above $3m, any property with idiosyncratic attributes or lease complexity, any property in a low-transaction suburb where the comparable set is thin, and any acquisition that will be financed by a lender who will require a certified valuation in any case.

The overlap between the two is the most productive zone. Using automated estimates to determine whether a certified appraisal is warranted, and then using the certified appraisal to calibrate the model's performance on that asset type, produces a feedback loop that improves screening accuracy over time.

The Confidence Interval Question

One operational problem with published AVM outputs is that they typically present a point estimate without a meaningful confidence interval. A single number without a band creates false precision. An automated estimate of $1.85m does not communicate whether the model is 90% confident the value sits between $1.8m and $1.9m, or whether it is genuinely uncertain across a $400,000 range.

We present our automated valuation outputs with an explicit confidence band tied to the comparability score of the property. A high-comparability property gets a tight band; a low-comparability property gets a wide band with a recommendation to seek a certified assessment before committing. That framing matches how acquisition professionals actually want to use the tool: with calibrated confidence, not manufactured precision.

The teams we have seen misuse automated valuation most consistently are those who treat the point estimate as the offer price rather than as the midpoint of a range that requires human judgment to narrow.

Calibrating Automated Models Against Your Own Deal History

Acquisition teams with an existing transaction history have a calibration resource that generic AVM products do not use: their own past deals. Properties that were acquired at specific prices and subsequently valued at different market levels provide a direct test of model accuracy within the specific asset class, suburb cluster, and price band that the team targets.

Running your acquisition history against the automated model retrospectively, asking what the model would have estimated for each asset at the time of acquisition, tells you where the model is systematically biased for your particular deal profile. If the model consistently undervalues commercial strata with short leases in your target suburbs by 12 to 18 percent, you can apply a calibration adjustment to that segment rather than treating the raw output as accurate.

This calibration process also reveals which property subtypes are genuinely outside the model's reliable range for your use case. For those types, the honest response is to route them directly to a certified appraisal without expecting the automated estimate to be useful. That segmentation, automated for high-comparability assets, certified for low-comparability ones, produces a higher-quality overall valuation process than applying the same tool uniformly to everything.